MDTL-NET: Computer-generated image detection based on multi-scale deep texture learning

Distinguishing between computer-generated (CG) and natural photographic (PG) images is of great importance to verify the authenticity and originality of digital images. However, the recent cutting-edge generation methods enable high qualities of synthesis in CG images, which makes this challenging task even trickier. To address this issue, a novel multi-scale deep texture learning neural network coined as MDTL-NET is proposed for CG image detection. We first utilize a global texture representation module incorporating the ResNet architecture to capture multi-scale texture patterns. Then, a deep texture enhancement module based on a semantic segmentation map guided affine transformation operation is designed for texture difference amplification. To enhance performance, we equip the MDTL-NET with channel and spatial attention mechanisms, which refines intermediate features and facilitates trace exploration in different domains. Moreover, a Low-rank Tensor Representation (LTR) strategy is also used for feature fusion. Extensive experiments on three public datasets and a newly constructed dataset1 with more realistic and diverse images show that the proposed approach outperforms existing methods in the field by a clear margin. Besides, results also demonstrate the detection robustness and generalization ability of the proposed approach to postprocessing operations.